Split Data Between Endpoints for Realtime Interference MACHINE-LEARNING-ASSOCIATE Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   What is data skew in the context of real-time inference endpoints?

   • AA situation where the training data is not shuffled before training
   • BA condition where the endpoint receives too many requests at once, causing latency spikes
   • CA situation where the model's predictions are consistently biased toward one class
   • DA mismatch between the distribution of data used to train the model and the data the endpoint receives
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A retail company deploys a demand-forecasting model to a real-time endpoint. The endpoint receives data from multiple stores. The team notices that stores in different regions have different data distributions. They want to ensure the model performs well for all regions. What is the most appropriate configuration?

   • AIncrease the endpoint's compute resources to handle all regions.
   • BRoute all traffic to the region with the highest data volume.
   • CDeploy separate endpoints for each region, each with a model trained on that region's data.
   • DUse a single endpoint with a model trained on all regions' data.
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A logistics company deploys a delivery-time prediction model to a real-time endpoint. The training data includes a categorical feature 'weather_condition' with values: sunny, rainy, snowy. At serving time, the endpoint receives a new value 'hailstorm'. The model returns an error. What is the most appropriate configuration to handle this?

   • AConfigure the endpoint to reject requests with unseen categorical values.
   • BConvert the categorical feature to a numerical one-hot encoding at serving time only.
   • CRetrain the model immediately whenever a new category appears.
   • DMap unseen categorical values to a designated 'unknown' category in the serving pipeline.
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A healthcare startup trains a patient-readmission model on historical EHR data. They deploy it to a real-time endpoint that receives data from hospital systems. The endpoint sometimes receives missing values for a feature that was always present in training. What is the best way to handle this in the serving pipeline?

   • AFill missing values with zero.
   • BRetrain the model to handle missing values.
   • CReject requests with missing values to ensure data quality.
   • DImpute missing values using the mean from the training data.
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A telecom company deploys a customer-churn model to two endpoints: one for residential customers and one for business customers. The residential endpoint receives 80% of the traffic. The team notices the business endpoint's accuracy is lower. They want to improve it without increasing costs significantly. What is the most appropriate action?

   • ARoute all business traffic to the residential endpoint.
   • BCollect more business customer data and retrain the model with a balanced dataset.
   • CIncrease the business endpoint's compute resources.
   • DReduce the business endpoint's request timeout to lower latency.
   Reveal Solution Discussion Select an answer first